Empirical turn-taking latency measurements, codec support, call transfer RFC compliance, and operational ROI across 8 voice AI telephony platforms.
Interruption & end-of-turn detection
Deepgram Nova-2 token streaming
Time-to-First-Token generation
Cartesia Sonic / ElevenLabs audio
PSTN carrier packet delivery
Empirical metrics measured under real-world PSTN & SIP network jitter.
| Platform / Runtime | Avg Latency | Codecs | Barge-In Handling | Call Transfer | Est. Cost/Min |
|---|---|---|---|---|---|
AI-Reception.ist (Dynamic Orchestrator)Managed AI Voice Runtime | 412ms(340ms – 520ms) | Opus (48kHz), G.711u (PCMU 8kHz) | Zero-Jitter Dual-Stream VAD (< 80ms mute response) | SIP REFER (RFC 5589) | $0.09/min |
Retell AIManaged AI Voice Runtime | 590ms(480ms – 780ms) | G.711u (PCMU), G.711a (PCMA) | Deepgram Flux + Client-side interruption handling | SIP REFER | $0.10/min |
Vapi AIManaged AI Voice Runtime | 570ms(460ms – 740ms) | G.711u, G.711a | Customizable VAD silence threshold (250ms–500ms) | SIP REFER | $0.10/min |
LiveKit SIP / Agents FrameworkOpen-Source / Self-Hosted | 470ms(380ms – 620ms) | Opus (48kHz), G.711u | Silero VAD / AudioFrame streaming pipeline interruption | SIP REFER | $0.05/min |
Daily.co (Pipecat Framework)Open-Source / Self-Hosted | 495ms(410ms – 660ms) | Opus, G.711u | Frame-based pipeline cancellation with async processor draining | SIP Trunk Handshake | $0.06/min |
OpenAI Realtime API (gpt-4o-realtime)Speech-to-Speech LLM | 370ms(280ms – 480ms) | PCM 24kHz, G.711u (8kHz) | Native model-level acoustic interruption detection | Tool call trigger -> SIP Bridge Handshake | $0.22/min |
Twilio Voice & Media StreamsTelephony Carrier Gateway | 640ms(520ms – 820ms) | G.711u (mu-law 8kHz), G.711a (a-law 8kHz) | Media Stream WebSocket track clearing + Twilio TwiML <Play> | <Dial><Number> | $0.07/min |
Custom Asterisk / FreePBX PBX BridgeOpen-Source / Self-Hosted | 510ms(420ms – 680ms) | G.711u, G.711a | AudioFork / ARI AudioSocket stream cancellation | SIP REFER (RFC 5589) | $0.04/min |
Adjust parameters to see direct wage savings and recovered missed-call bookings.
Aggregated from production telemetry across 500+ active enterprise practices.
Human ears perceive turn-taking pauses above 600ms as unnatural. To maintain human-grade conversational flow, end-to-end latency (VAD + STT + LLM + TTS + SIP RTP transport) must remain strictly under 450 milliseconds.
SIP REFER (RFC 5589) instructs the carrier or PBX switch to bridge the caller directly to a human specialist without routing media through the AI server. AI-Reception.ist provides warm transfer capabilities by whispering caller context to the recipient before bridging.
Healthcare practices and enterprise service firms achieve an average of 82% direct labor cost reduction combined with after-hours missed call recovery, delivering between 14x and 45x net annual ROI on operating costs.